The software path
Applicationyour workload and framework
Librariesfor example cuDNN or TensorRT
Runtime + toolkitAPIs, compiler, development tools
Driveroperating-system interface and CUDA capability
HardwareGPU features and compute capability
A version printed at one layer cannot by itself establish the version or support state of every other layer.
[5][32][34][35]The physical integration path
GPU siliconarchitecture implementation
Module or boardmemory, power, form factor
ServerCPU, storage, local fabric
Rack-scale systemmultiple compute and switch trays
Network + facilitycluster fabric, power, cooling
Use the map to diagnose the claim
If a press release says a platform is “available,” ask: which GPU or system, through which provider or channel, in which region, with which software version? If a workload fails, ask: is the issue hardware feature support, driver capability, toolkit installation, a library matrix, or application packaging?
The map turns a vague product question into a sequence of checkable claims.
[23][32][31]